topic Re: Apparent discrepancy between GEE estimation and Type-3 analysis in PROC GENMOD in Statistical Procedures

Apparent discrepancy between GEE estimation and Type-3 analysis in PROC GENMOD

ABiostatistician — Wed, 19 Aug 2015 20:30:49 GMT

Dear Community,

I have a data set with a binomial response and a binary predictor, q. My observations are clustered under the various values of a categorical variable, r. I ran this following code:

proc genmod

data = fakedata2;

class

q (ref = '0')

model successes/trials

/ dist = bin

link = logit

type3;

repeated

subject = r;

run;

My GEE estimate for Q = 1 has a P-value of 0.0008, but the Type-3 score statistic for Q has a P-value of 0.0956.

My questions for you:

1) How is this possible? (I've read the documentation for Type-3 analysis, and I can't figure out how this is exactly calculated. I admit that I don't fully understand what the Type-3 analysis is doing, but, as I have learned from other statisticians, the Type-3 analysis should test for the overall significance of Q while controlling for other effects. Since Q is my only predictor, I would expect their P-values to be similar, if not the same.

2) Given these disparate P-values, what can I conclude about Q from this model? Is Q a significant predictor of success? Why or why not?

Thanks for your help.

Re: Apparent discrepancy between GEE estimation and Type-3 analysis in PROC GENMOD

SteveDenham — Fri, 21 Aug 2015 13:18:32 GMT

The type III tests are for a model that assumes that the individual time points are independent (no correlation, and from independent observations as well). The solution serves as the initial values for the GEE, which models the correlation between time points. Here, you model the default exchangeable correlation structure. It is not at all surprising that the two p values differ substantially, if there is a correlation between time points within subject. That is what a GEE approach is designed to handle.

Steve Denham

Re: Apparent discrepancy between GEE estimation and Type-3 analysis in PROC GENMOD

ABiostatistician — Mon, 24 Aug 2015 22:44:11 GMT

Thanks for your helpful reply, Steve.

If my Type-3 test shows a high P-value, but my GEE estimate shows a low P-value, then which one should I trust?

I see this type of discrepancy in 2 situations:

a) there is only 1 binary predictor, Q

b) there are 3 categorical predictors: Q (binary), X (ternary) and W (ternary).

Thanks for your insights!

Re: Apparent discrepancy between GEE estimation and Type-3 analysis in PROC GENMOD

ABiostatistician — Thu, 27 Aug 2015 03:29:39 GMT

Hi Steve,

I finally learned the reason. If I invoke the "WALD" option in the "MODEL" statement, then both the GEE and the Type-3 analysis give the same P-values.

By default, PROC GENMOD uses score tests for Type-3 analyses, and that resulted in the difference in P-values in my model.